A clot-risk model looked strong once, then weaker across repeated splits
In 380 Chinese hip-arthroplasty patients, a random forest reached 0.859 AUC on one held-out group but averaged 0.792 across 30 repeat partitions. Two traditional-medicine features added no measurable discrimination.
The short version
- What happened
- The single-centre retrospective cohort contained 380 patients and 105 postoperative DVT events; the held-out group had 113 patients and 31 events.
- Who it affects
- A hip-fracture patient needs rapid clot prevention and investigation based on the whole clinical picture. A model that performs differently across internal splits should not decide who receives either.
- What remains unproven
- 380-patient retrospective cohort
Keep in mind
380-patient retrospective cohort
Single centre
105 total DVT events
31 held-out events
Internal validation only
No prospective workflow
Funding and conflicts not yet exposed in the accepted abstract
Source analysis · 1 linked source. Primary analysed source date: 29 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The repeat tests matter more than the best-looking split
A random-forest model identified most postoperative deep-vein thromboses in one held-out group of hip-fracture patients, but its performance fell when researchers repeated the partitioning, according to a peer-reviewed Frontiers in Artificial Intelligence paper published on 29 October. The selected split produced an AUC of 0.859 and sensitivity of 90.3%. Across 30 repeated constrained partitions, however, mean validation AUC was 0.792—a more sober estimate of how unstable a small internal dataset can be.
For a surgical team, the possible use is an extra risk prompt before or after hip arthroplasty, not a decision to withhold ultrasound, thromboprophylaxis or clinical review. At the chosen threshold, specificity was 72.0%, positive predictive value 54.9% and negative predictive value 95.2%. Those values came from a cohort in which 27.6% developed DVT; they will change in hospitals with a different patient mix and screening practice.[1]
Eight variables, nine algorithms and 105 events
The retrospective single-centre study included 380 people undergoing hip arthroplasty for femoral-neck fracture in China; 105 developed lower-limb DVT. Researchers placed 267 patients with 74 events in development and 113 patients with 31 events in a profile-disjoint held-out group. Identical prespecified multivariable profiles were kept from crossing evaluation boundaries. They compared nine machine-learning algorithms using eight fixed preoperative variables, then selected random forest on development-set out-of-fold performance before looking at the holdout.
The variables were D-dimer, fibrinogen, platelet count, type 2 diabetes, fasting glucose, the triglyceride-glucose index and two traditional Chinese medicine syndrome classifications. Development out-of-fold AUC was 0.774. Calibration in the single holdout had a slope of 1.136 with a wide 95% confidence interval from 0.753 to 1.867, an observed-to-expected ratio of 0.846 and a Brier score of 0.132. With only 31 held-out events, those estimates remain imprecise.[1]
The distinction that matters
One favourable split versus repeated internal tests
The headline metric changes when the same dataset is partitioned more than once.
Development
Group-aware out-of-fold testing
AUC 0.774
Selected holdout
113 patients; 31 DVT events
AUC 0.859
30 repeat partitions
Repeat constrained validation
Mean AUC 0.792
The traditional-medicine variables did not improve ranking
Adding the two retained traditional-medicine syndrome features changed the holdout AUC from 0.855 to 0.859, a difference of 0.003 that was not statistically significant (DeLong p=0.746). By contrast, a one-standard-deviation increase in the triglyceride-glucose index remained associated with DVT after adjustment (odds ratio 1.55, 95% CI 1.186–2.026). Association does not prove that lowering that index would prevent a clot, nor that it adds value beyond an established clinical pathway.
The model has not crossed the main barrier to clinical use: testing in a separate hospital and then prospectively in care. The accessible accepted abstract does not yet expose the final funding and competing-interest statements, so those details also require checking when the formatted article appears. The assessment would change with external calibration, comparison against a simple clinical score, subgroup error analysis and evidence that use improves timely detection without unnecessary scans or treatment.[1]